Full text: Proceedings of the Symposium on Global and Environmental Monitoring (Part 1)

the total line length (3 x 25 m = 75 m) is an 
estimate of the relative cover of the cover type 
within the ground element. These relative 
covers per ground element serve as the X, in 
the regression analysis. 
MICROTINO TRANSECT LOCALIZATION Al.FA - 7M.30 
Contours for RESTVAR 
Figure 4: Contours of residual variance as a 
function of the array position. The graph 
occupies an area of 1 pixel (25 x 25 m), illu 
strating the subpixel accuracy. Contour values 
are relative. 
EVALUATION 
The predictions were evaluated by calculating 
the root mean squared error of prediction per 
cover type k (RMSEP k ), defined as: 
RMSEP. = (n ’ S ix,, - x",) 2 )'' 2 (4) 
I = 1 
where n is the number of evaluation ground 
elements and x jk and x ik are observed and 
predicted cover fractions, respectively. 
In this case there was no seperate test set 
available, so the training sets were also used 
for the evaluation. Instead of resubstituting all 
training pixels in (4), a lcave-one-out (LOO) 
method was used to calculate RMSEP. In 
LOO, regression coefficients are calculated 
using a training set of n-1 pixels, after which 
the pixel that was left out is predicted. This 
procedure is repeated n times for all training 
set pixels. LOO gives a more realistic estimate 
of prediction error, especially for small sample 
sizes. 
Cross predictions between training sets give 
an estimate of predictive power of regression 
models when applied in area’s other than that 
of the training set. 
BAND SELECTION AND PREPROCESSING 
It has been proposed to select the combi 
nation of bands that has the largest data vari 
ance (Sheffield 1985). Here, the capacity of ggg 
several band combinations to explain the varia 
tion in heather and grass cover was used as 
critcrium for band selection. Figure 5 shows 
residual variance of prediction for heather 
cover as a function of band combination. Only 
bands 3, 4, 5 and 7 were considered. Band 1 
and 2 were omitted because of the high noise 
level and the strong correlation with band 3 
respectively, while band 6 was left out because 
of the different origin of the recorded radia 
tion: active thermal instead of reflective. Sev 
eral combinations of the remaining bands 
proved to have a statistically equal explanatory 
capacity, all incorporating band 5. Because 
differences were small and only based on 
small data sets, all four bands (3, 4, 5, 7) 
were used in subsequent regression analyses. 
The aim of this study was not to model 
the functional relationship between reflectances 
and ground cover, but to calibrate the avail 
able Landsat scenes. For that reason, no 
attempt has been made to transform the digit 
al numbers to a physical unit with more gen 
eral validity and thus no further radiometric 
correction has been applied. As a result, the 
regression equations are only valid for this 
specific case. 
The images have been corrected geometric- 
ly to standard topological map projection to 
facilitate delineation of area’s of interest. By 
resampling, the pixclsize was reduced to 25 m. 
Figure 5: Residual variance of prediction for 
grass cover with several band combinations 
RESULTS 
Dutch healhland occupies about 42(XX) ha. 
Only part of this area consist of dry heath 
that formed the object of interest for this 
study. Leaving out wet heath, coastal heath, 
fens, roads, large patches of bare ground, 
woods and woodland, some 17(XX) ha was 
selected to be mapped. This area was covered 
by two Landsat quarter scenes, from July 16 
and August 4, 1986 respectively. Three arrays 
of ground elements were measured in the 
field, one in the north, one in the middle and 
one in the south of Holland, containing 95
	        
Waiting...

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.